Example 0717 now asserts the (token, index) Diag for ALL SIX raise sites
in cli.sx, closing the two the reviewer found still unasserted:
- zero-arg UnknownCommand: parse([], ...) -> index -1, token ""
(the args.len == 0 sub-branch of cli.sx:237, distinct from the
one-arg too-few form already covered at index 0 / token args[0]).
- TooManyFlags (cli.sx:256): a command declaring 17 flag specs (> the
inline 16 cap) is rejected, not truncated -> index -1, token command.
The three index==-1 cases (zero-arg, too-many, missing-req) seed their
Diag with a sentinel before parse, so each assertion proves parse WROTE
the -1/"" rather than merely matching the `.{}` default. Verified
non-vacuous: flipping any expected value makes that line FAIL.
Test-only: cli.sx logic and src/ are untouched.
Extend example 0717 to pin the offending token VIEW and its args index
for every failure the parser's Diag populates: unknown-command,
unknown-group, too-few-args, missing-value, value-eats-flag, and the
missing-required index. Closes the test-coverage gap flagged in review;
cli.sx parser logic unchanged.
Extend std/cli.sx with a zero-heap argument parser that the caller drives
over a logical argv ([]string), separate from the F3.1 os_args accessor.
Grammar: <group> <command> [--flag VALUE | --bool]... [--json] [-- rest...]
- (group, command) dispatched against a caller-provided Command table;
no match -> error.UnknownCommand.
- value-taking vs boolean flags fixed by each command's FlagSpec list;
--json is a reserved global boolean surfaced as parsed.json.
- `--` or the first bare operand ends flag parsing; the remainder is
parsed.rest (operand views).
Heap discipline (heap-discipline.md): zero heap, zero copy. group/command/
flag values/rest are all VIEWS into args. Parsed is a by-value stack struct;
flag presence/values live in a fixed [16]FlagValue inline array indexed by
spec position (no per-flag allocation, no context.allocator). The flag-spec
list and command table are caller storage passed as views.
Failure surfacing (no silent skip): unknown command, unknown flag, a
value-flag missing its value, and an absent required flag each raise a
specific CliError variant; a caller-owned Diag records the offending token
(index + view) before each raise, since error tags carry no data.
examples/0717 drives the parser over explicit []string vectors: a valid
group/command/--flag/--bool/--json case (asserting parsed values + that
values are views into argv), subcommand dispatch, `--`/bare-operand
separators, and the five failure variants each asserted via destructure +
Diag. zig build && zig build test && run_examples.sh green (385 passed).
The global-init constant serializers in emit_llvm.zig printed a diagnostic
on an unserializable value and then RETURNED an undef/null placeholder and
CONTINUED emitting. For a comptime `#run` global that yields a function
reference (`fp :: #run pick();` where pick returns a function), the build
fell through to the JIT and segfaulted calling through the undef pointer
(exit 134) — a silent miscompile dressed up as a printed error.
Route every genuine bail in the serialization family through a new
`failGlobalInit` helper: it sets `comptime_failed` (so core.generateCode
aborts with a non-zero exit after emit()) and returns an undef placeholder
that never ships, because the halt fires before object emission / JIT. This
covers the comptime func_ref leaf, the require_resolved aggregate func_ref
leaf, the top-level + vtable func_ref globals, the comptime-init catch, and
the remaining heap-walk / aggregate-shape bails. Unresolved-function
diagnostics now name the function instead of its (stdlib-unstable) IR index.
The require_resolved=false Pass-0 placeholder is unchanged (func_map is
empty until Pass 1; the aggregate is re-emitted with require_resolved=true).
Regression: examples/1128-diagnostics-comptime-global-funcref-rejected.sx —
a `#run` global returning a function ref now exits 1 with the diagnostic
(was: exit 134 segfault). Fail-before/pass-after verified.
A module-global initialized with an enum literal silently zero-initialized
to the first tag (`chosen : Color = .green` read back as `.red`), and an
enum tag inside a global array/struct was rejected as non-constant. The
constant serializer had no enum-literal arm.
Add `Lowering.constEnumLiteral`: serialize an enum literal to a
`ConstantValue.int` holding the variant's tag value, resolved against the
destination enum type and respecting explicit variant values; the global's
type drives the backing width at emit time. Wired into `globalInitValue`
(scalar global) and `constExprValue` (array element / struct field / nested
aggregate). A non-enum destination or unknown variant is diagnosed loudly,
never silently zero-initialized. The compiler-injected OS/ARCH globals now
serialize to their real `.unknown` tag (6 / 4); runtime reads are unchanged
(they resolve through comptime_constants), so only the static initializer in
the pinned .ir snapshots changes.
Remove the silent `func_ref => orelse LLVMConstNull` fallbacks in the LLVM
constant emitters: aggregate func_ref leaves carry a `require_resolved` flag
(transient null in Pass 0, loud diagnostic if still unresolved in the
Pass-1.5 re-emit), a top-level func_ref global is resolved in
initVtableGlobals, and the comptime (#run) path bails loudly instead of
emitting a null function pointer.
Regression: examples/0139-types-global-enum-literal-init.sx (scalar, array,
struct field, explicit-value enum u16 stride, struct-array with enum field);
negative: examples/1127-diagnostics-global-enum-literal-bad-variant.sx.
Mark issue 0082 RESOLVED.
A module-global aggregate initializer rejected a `null` literal in a
pointer (or optional-pointer) field as "must be initialized by a
compile-time constant". `Lowering.constExprValue` had no `.null_literal`
arm, so the null leaf returned no constant and the whole aggregate looked
non-constant — even though `null` is the compile-time zero pointer (a
top-level scalar `p : *s64 = null;` already serialized fine).
Add `.null_literal => .null_val` to constExprValue. While here, make the
two LLVM constant emitters exhaustive: emitConstAggregate and the
top-level init_val switch in emit_llvm.zig previously ended in a silent
`else => LLVMConstNull(...)` catch-all (the silent-arm class CLAUDE.md
mandates rooting out). They now handle every ConstantValue tag explicitly
(.null_val/.zeroinit -> all-zero constant, .undef -> LLVMGetUndef,
.func_ref resolved, nested .vtable is a hard @panic tripwire). The
reject-loud path for genuinely non-constant fields is preserved.
Regression: examples/0138 (array-of-struct null ptr fields, array of
all-null pointers, nested struct-in-struct null ptr) and the negative
examples/1126 (null ptr field beside a non-const field still errors).
Fail-before/pass-after verified.
A module-global array of struct literals (`pairs : [2]Pair = .[ .{...}, .{...} ]`)
was emitted as `zeroinitializer`, silently dropping every declared field — reads
returned 0 with no diagnostic. Global struct literals and struct-with-array
already worked; the gap was struct literals used as ARRAY elements.
Root cause: `Lowering.constExprValue` (the const-aggregate serializer for global
initializers) had no `.struct_literal` arm. `constArrayLiteral` serialized each
element through `constExprValue`, so a struct-literal element returned null,
collapsing the whole array initializer to null; `globalInitValue` then emitted no
payload and the LLVM backend zero-initialized the global — the same silent-zero
class as 0071/0072, one level inside an array literal.
Fix: make `constExprValue` type-aware — thread the destination element/field
TypeId so a struct-literal leaf routes through `constStructLiteral` and a nested
array-literal through `constArrayLiteral` with the correct element type.
`constArrayLiteral` derives its element type from the array TypeId;
`constStructLiteral` passes each field's type. A global aggregate initializer that
still does not fully reduce to a compile-time constant is now rejected loudly
(`diagnoseNonConstGlobal`) instead of silently zeroing. `emitConstAggregate`
already recurses over nested aggregates, so `sx run` (JIT) and `sx build` (AOT)
both materialize the declared values.
Regression: examples/0137-types-global-aggregate-literal-init.sx (global
[N]Struct literal, global struct literal, struct-with-array, nested
array-of-struct-with-array; values read back with no prior store, plus a store on
top). Fails on the pre-fix compiler (array-of-struct fields read 0), passes after.
Marks issues 0079 (already resolved) and 0080 RESOLVED.
A store to a module-global array element (`g[i] = v`) was silently dropped:
a subsequent `g[i]` read the array's initializer, not `v`. Constant index,
variable index, and cross-function stores were all affected, in both `sx run`
and `sx build`. Global scalars and local arrays were fine.
Root cause: `Lowering.lowerExprAsPtr` (the lvalue/address path) handled only
local identifiers. A module-global identifier fell through to the value
fallback `lowerExpr`, which emits `global_get` — loading the whole array by
value. The LLVM backend's `emitIndexGep` then allocas a throwaway temp, copies
the value in, and GEPs into the temp, so the store wrote a discarded copy.
Fix: teach `lowerExprAsPtr`'s identifier arm about globals — emit `global_addr`
(a pointer into the global's live storage), or `global_get` for a pointer-typed
global (mirroring the local pointer case). Route the `address_of(index_expr)`
array base through `lowerExprAsPtr` too so `&g[i]` is likewise an lvalue into
the global. `index_gep` now GEPs directly into the global for const and variable
index, across functions. This also fixes global struct field stores, which
shared the same root cause.
Regression: examples/0136-types-global-array-element-store.sx (const-index,
var-index, cross-function store on a scalar global array; struct-element array
for stride; nested-array global for the recursive lvalue). Fails on the pre-fix
compiler, passes after.
Add library/modules/std/cli.sx: a pure-sx command-line argument accessor
backed by the macOS C runtime (_NSGetArgv/_NSGetArgc), no compiler change.
os_argc() -> s64
os_args(buf: []string) -> []string
Zero heap, zero per-arg allocation: os_args fills a caller-provided buffer
(stack array) with string VIEWS over the process's own argv block, which
lives for the whole process. The returned slice header is a by-value stack
return; nothing touches context.allocator.
Documents the `sx run` reality: under `sx run <prog.sx> ...` the process
argv is the interpreter's argv (sx, run, prog.sx, ...), not a program's
logical args. This accessor reports the real process argv truthfully;
mapping to logical args is a later consumer concern (distribution P3.1).
Non-macOS platforms bail loudly (message + _exit) rather than returning a
silent empty.
examples/0716-modules-cli-argv.sx asserts only deterministic structural
invariants (argc >= 1, argv[0] non-empty, os_argc() == filled length).
Add 0715-modules-json-suite as the single comprehensive pinned suite for
std.json (mirrors 0711 for std.hash), alongside the focused 0713/0714 demos:
- ROUND-TRIP build->write->parse->write over a document covering EVERY value
kind (a string with every escape form \" \\ \b \f \n \r \t plus a \u00XX
control, integers 0 / negative / s64 MIN / s64 MAX, bool, null, array,
nested object) with insertion-order assertions, exact writer bytes, and
parse-then-rewrite idempotence.
- DECODE positives: \/, the full named-escape set, \uXXXX (BMP 1- and 2-byte)
plus a surrogate pair, the escaped control forms, and raw multi-byte UTF-8
round-tripping through writer + reader.
- MALFORMED matrix: one assertion per JsonParseError variant and its key
edges (UnexpectedToken, UnexpectedEnd, BadEscape, BadNumber incl. leading
zero / lone '-' / fraction / exponent / overflow, TrailingGarbage,
BadControlChar), each asserted to raise.
Pure test work: src/ and library/ untouched, no json.sx change needed. Every
model is built through an explicit Arena allocator (heap discipline).
parse_string scanned for `"` and `\` but accepted every other byte,
including raw control characters. RFC 8259 §7 requires those bytes to be
escaped inside a string; an unescaped one is invalid JSON and must surface
a parse error, not be silently accepted.
Add `BadControlChar` to JsonParseError and reject any unescaped byte < 0x20
in the string body scan (which gates the decode path too, so escaped forms
like \t/\n/ still decode correctly; 0x20 and 0x7F are not over-rejected).
Regression test in examples/0714: raw 0x09/0x0A/0x00 each raise
BadControlChar via `?`/`!`; a positive case proves the escaped forms still
decode to the right bytes. All prior assertions kept.
Issue 0078 (string == as an and/or operand emitting an invalid PHI) is
resolved on this branch, so the example no longer needs the split that
worked around it. Restore the natural combined assertion
sub.items[0].key == "k" and sub.items[0].val.str == "v"
(one nested-pair report), and the in_range containment helper to
return x >= lo and x < hi;
Drop the now-stale issues/0078 references. Re-captured expected stdout
(nested-key/nested-val -> nested-pair). json.sx and src/ untouched.
A string `==`/`!=` used as an operand of a short-circuit `and`/`or` emitted
invalid LLVM (`PHI node entries do not match predecessors!`). String compares
expand into their own memcmp sub-CFG during LLVM emission, so the operand
finishes in a later basic block (`str.merge`) than the one the IR block
started in. `fixupPhiNodes` wired the short-circuit merge PHI's incoming edge
to `block_map[ir_block]` (the block the IR block started as), recording a
stale predecessor (`%entry`/`%and.rhs.0`).
Fix: record the builder's actual insertion block after emitting each IR
block's instructions (`term_block_map`, via `LLVMGetInsertBlock`) and use it
as the PHI predecessor. General — corrects the incoming block for any operand
that emitted intermediate basic blocks (string `==`, value `match`, …), not
just string `==`.
Regression: examples/0045-basic-string-eq-short-circuit.sx (string `==` on
both sides of `and` and of `or`, plus a match-value + enum-payload `==` shape).
Fails (LLVM abort) pre-fix, passes after.
Add the JSON reader (parser) to library/modules/std/json.sx, the inverse
of the F2.1 writer over the same value model: insertion-ordered objects,
arrays, strings (full unescaping incl. \uXXXX + surrogate pairs), s64
integers, bool, null.
Heap discipline (binding): exactly two allocation kinds, both through the
EXPLICIT `alloc` parameter, never the implicit context allocator —
composite backing stores (Array/Object.items via add/put) and decoded
escaped-string buffers (bounded by the raw span). Un-escaped string
values are zero-copy VIEWS into the input buffer (valid only while it
lives); scalars carry no heap.
Failure surfacing (hard contract): malformed input raises a meaningful
JsonParseError variant (UnexpectedToken / UnexpectedEnd / BadEscape /
BadNumber / TrailingGarbage) on the error channel, never a bogus value.
Trailing non-whitespace is TrailingGarbage; fractions/exponents,
out-of-s64 magnitudes, and leading zeros are BadNumber. Number
accumulation runs in negative space so s64 MIN parses exactly.
examples/0714-modules-json-reader.sx asserts the parsed structure
(insertion order, every kind), proves the view-vs-decoded heap split by
pointer containment, round-trips back through the writer byte-for-byte,
decodes a surrogate-pair into 4 UTF-8 bytes, and checks every malformed
variant.
Filed issues/0078: a string `==` (or any sub-CFG operand) used in a
short-circuit `and`/`or` emits invalid LLVM IR (stale PHI predecessor),
hit while writing the example's assertions and worked around there by not
combining comparisons with `and`/`or`. src/ untouched.
Close the coverage gap from attempt 1: example 0713 now builds integer
fields holding s64 MIN (-9223372036854775808) and s64 MAX
(9223372036854775807) — plus zero, a small negative, and a small positive —
and asserts the EXACT emitted bytes. This permanently pins the edge that
write_int is specifically engineered for (folding positives into negative
space so MIN's non-representable-positive magnitude serializes correctly).
s64 MIN is expressed as (0 - 9223372036854775807 - 1) because its magnitude
is not a representable positive s64 literal.
Test hygiene: stream to a repo-local, gitignored .sx-tmp/ path (created if
missing) instead of a fixed /tmp name, and unlink it right after read-back
so nothing leaks. Writer/model logic and src/ are untouched.
Add library/modules/std/json.sx — the JSON value model and writer
(reader lands in a later step).
Value model: a tagged union over null/bool/integer(s64)/string/array/
object. Objects are an ORDERED list of (key,value) pairs preserving
INSERTION ORDER (no hash map, never sorted/deduped). Integers only — no
fraction/exponent this milestone.
Heap discipline:
- Scalars carry no heap; string values are VIEWS into caller memory
(never copied into the node).
- Composite nodes (Array/Object) own growable child storage, allocated
through an EXPLICIT allocator parameter on the builder methods
(arr.add(v, alloc) / obj.put(key, val, alloc), mirroring List.append)
— never the implicit context allocator.
- The writer adds ZERO output allocations: it emits into a caller-
provided Sink, either a fixed []u8 buffer (overflow raises, never
truncates) or streaming straight to an fs.File through a small caller
staging buffer (no whole-document string; peak memory O(staging)).
Integer digits format in a stack [20]u8; s64 MIN is handled by
formatting in negative space. Sink/IO/overflow surface on the !
error channel.
examples/0713-modules-json-writer.sx builds a nested object + array +
string with every escape kind + negative int + bool + null, then asserts
the EXACT bytes (insertion order, escaping) from both the buffer sink and
the file-streaming sink, plus the overflow-raises path.
Make the SHA-256 digest path allocation-free (foundation heap-discipline):
- final() and sha256_hex() now return the 64-char lowercase hex digest as
a [64]u8 by value on the stack; the cstring(64) heap allocation is gone.
- sha256_file() streams the file in fixed 64KB stack chunks via open_file/
File.read/File.close (defer-closed on every path) instead of slurping it
with read_file; peak memory is O(chunk), not O(filesize).
Tests (compare via a zero-copy string view over the [64]u8):
- 0710 updated to the by-value API (output unchanged).
- 0711 known-answer vectors: "", "abc", NIST-56/112, padding boundaries
{0,55,56,57,63,64,65,119,120}, and 1000 / 1,000,000 'a' repeats, each
pinned to its published digest (cross-checked with shasum -a 256).
- 0712 streaming equivalence (one-shot == byte-at-a-time == split-mid-block
== split-on-boundary) plus sha256_file(temp) == in-memory digest.
src/ untouched. zig build && zig build test && tests/run_examples.sh green.
Add a pure-sx streaming SHA-256 (FIPS 180-4) stdlib module, importable
as `#import "modules/std/hash.sx";`. All 32-bit word arithmetic is done
in s64 and masked back with `& MASK32`, so digests are deterministic and
platform-independent — no shelling out, no native crypto.
API:
- init() -> Sha256 (by-value *self pattern)
- update(*Sha256, string) (multi-block + partial-block buffering)
- final(*Sha256) -> string (32-byte digest as lowercase hex)
- sha256_hex(string) -> string (one-shot)
- sha256_file([:0]u8) -> ?string (digest of a file via fs.read_file)
Verified against FIPS/NIST known-answer vectors and `shasum -a 256`:
"" , "abc", the 56- and 112-byte multi-block vectors, 1000×'a', and the
64/65-byte block boundaries; chunked update() matches the one-shot call.
examples/0710-modules-sha256.sx pins the KAT vectors + the streaming
invariant; gate green (zig build, zig build test, run_examples 370/0/0/0).
The reserved-type-name binding diagnostic fired correctly but underlined the
enclosing statement / if / while / for / match / protocol / #objc_class block
because every binding-name check reused the parent `node.span`.
Thread each binding name's own span through the AST and parser, and pass it to
`checkBindingNames`:
- ast: add name spans to VarDecl, DestructureDecl, If/WhileExpr, ForExpr
(capture + index), MatchArm, Catch/OnFailStmt, Protocol/ForeignMethodDecl.
- parser: populate each span at the binding site from the name token's loc;
destructure reuses each target identifier's own span.
- semantic_diagnostics: every checkBindingName call now passes the binding's
own span — no site falls back to node.span. fn/lambda params already used
Param.name_span.
Carets now land on the offending identifier itself. New regression
examples/1125 asserts the protocol default-body and sx-defined #objc_class
method param spans; 0125/1119-1124 expected updated to the precise carets.
The reserved/builtin-type-name binding diagnostic was a hand-walked subset
of binding-bearing AST nodes with a silent `else => {}`, so each review
found another syntactic binding form that bypassed it and hit the original
LLVM verifier abort: destructure names (`s2, x := …`), `impl` method
params/locals, and `if` / `while` / `for` / match-arm / `catch` / `onfail`
captures.
Rewrite `checkBindingNames` (src/ir/semantic_diagnostics.zig) as an
EXHAUSTIVE `switch` over every `Node.Data` tag with NO `else` arm — a future
binding-bearing node type now fails to compile until it is handled here, so
coverage is enforced by the compiler instead of a hand-maintained list. The
check stays in the pre-lowering semantic pass rather than moving to the
`Scope.put` scope-registration choke point: lowering is lazy, so an
uncalled function's bindings never reach `Scope.put`, yet they must still be
rejected at their declaration (e.g. the never-called `takes_u8` in 1119).
No lowering special-case; `lower.zig` unchanged.
Regression tests (fail-before: LLVM abort or silent accept → pass-after:
clean diagnostic, exit 1):
- 1121 control-flow: destructure, if/while bindings, for capture+index,
match-arm capture
- 1122 impl-block method: reserved param AND reserved local
- 1123 catch + onfail tag bindings
- 1124 destructure name reserved in an imported module
Existing 0125 / 1119 / 0135 / 1120 tests kept; full suite 368 passed.
The issue-0076 reserved-type-name binding diagnostic only ran over main-file
decls, so an imported module (or the stdlib) could still declare `s2 := ...`
and reach lowering, where the address-of family loads the whole aggregate and
passes it by value to a `ptr` param — LLVM verifier abort.
Extend coverage to every compiled module: a dedicated `checkBindingNames` walk
(in semantic_diagnostics.zig) visits every var/`:=`/typed-local binding name and
function/lambda/struct-method parameter at any depth, with NO main-file filter,
descending the `namespace_decl` that a `mod :: #import` wraps so imported-module
decls are reached. It tracks each module's source_file (save/restore per node)
so the diagnostic renders against the imported module's text. Rejection still
defers to the parser's `Type.fromName` classifier; the unknown-type check (0064)
stays main-file-only. No lowering special-case; `.identifier`-only address-of
paths are unchanged.
Stdlib audit: the only reserved-name bindings under library/ were two `u1`
locals in ui/renderer.sx (UV coords) — renamed to u_min/u_max/v_min/v_max.
Regression test: examples/1120-diagnostics-imported-reserved-type-name.sx (+
companion mod.sx) — an imported `s2 := ...` now emits the clean diagnostic at
the import's declaration site (exit 1), not an LLVM abort.
Resolves issues 0076 (coverage extension) and 0077.
A value binding (local/global `var` or a parameter) spelled as a
reserved/builtin type name parses as a `.type_expr` rather than an
`.identifier` (parser.zig, via `Type.fromName`), so the address-of
family in lower.zig never saw a scoped local and mis-lowered it —
loading the aggregate and passing it by value to a `ptr` parameter
(LLVM verifier abort, or a silent `*self`-mutation-losing copy).
Add a declaration-site diagnostic in semantic_diagnostics.zig
(`UnknownTypeChecker.checkBindingName`): reject any parameter name or
`var` binding name (`:=` / typed-local / global forms) whose spelling
collides with a reserved type name. `isReservedTypeName` defers to the
parser's own classifier (`types.Type.fromName`) so the rejected set
never drifts from the set that would parse as a type — the named
builtins (bool/string/void/f32/f64/usize/isize/Any) and `[su]N` over
sx's 1-64 range. Bare value names (`s`, `self`, `index`) are untouched.
No lowering special-case; the `.identifier`-only address-of paths are
correct once type-shaped names can never be bound. The rejected
attempt-1 `bareVarName` approach was never landed.
Tests:
- 0125-types-type-named-var-rejected: `:=` form (s2) rejected
(repurposed from the old test that asserted the now-illegal behavior).
- 1119-diagnostics-reserved-type-name-as-identifier: parameter (u8),
typed-local (s64, bool), `:=` (string) forms rejected.
- 0135-types-self-streaming-nonreserved: positive — `*self` streaming
with non-reserved names accumulates correctly via both call styles.
- 0904-optionals: renamed incidental locals s1/s2 -> filled/empty.
Backend-FFI .ir inventory + scaffolding for the Obj-C/JNI runtime-constructor
extraction (Phase A7.3). No code moved.
Inventory (recorded in ARCH-SAFETY.md): the existing FFI .ir set already pins the
core constructor emission — emitObjcSelectorInit (sel_registerName via 1309/1329/
1332), emitObjcClassInit (objc_getClass), emitObjcDefinedClassInit class
registration + ivars + method IMP table (objc_allocateClassPair / class_addIvar /
class_addMethod / objc_registerClassPair via 1309/1332), and getOrCreateJniSlots /
emitJniConstructor (GetMethodID via 1402/1418/1408).
Gaps closed (2 new .ir snapshots) for the ARCH-SAFETY-named metadata not covered
by 1309:
- 1319-ffi-objc-property-sx-defined: property getter/setter IMPs (_get/_set/
class_addMethod x8).
- 1314-ffi-objc-class-dealloc-roundtrip: alloc/dealloc IMPs.
Both path-free + idempotent (verified across two captures; trailing newline
trimmed). Suite count unchanged (snapshots on existing examples).
Gate: zig build, zig build test, bash tests/run_examples.sh -> 361/0 (no churn
beyond the 2 new .ir).
Codex review of d6078c2 flagged a blank line at EOF in the new
examples/expected/1202-ffi-cc-c-large-aggregate.ir. Collapse the trailing
newlines to a single one so `git diff --check` is clean. Test-safe: the runner
reads both expected and actual IR through $(...) command substitution, which
strips trailing newlines, so the comparison is unaffected (1202 still ok).
Gate: zig build, zig build test, bash tests/run_examples.sh -> 361/0.
Test-first scaffolding for LLVM backend modularization (Phase A7.1) before the
type/ABI helpers move into src/backend/llvm/{types,abi}.zig. Visibility-only
change to the targets — no behavior change. Closes the ARCH-SAFETY "no generic
ABI snapshot" gap.
- 2 new emit_llvm.test.zig tests:
- abiCoerceParamType across every C-ABI size bucket: <=8 -> i64, 9-16 ->
[2 x i64], >16 -> ptr, HFA (all-float/all-double, <=4 fields) -> unchanged,
string -> ptr, slice -> ptr, scalar -> unchanged. Built via a local
internStruct helper (field slice in the module arena -> no testing-allocator
leak); asserts against emitter.cached_* + LLVMArrayType2.
- needsByval: true only for >16-byte non-HFA struct; false for <=16 / HFA /
string / slice / non-struct.
- 1 new .ir snapshot: 1202-ffi-cc-c-large-aggregate (the canonical callconv(.c)
>16-byte byval example that directly documents abiCoerceParamType) — pins the
byval param path end-to-end (5 byval + entry reload + 2 sret from Arena.init).
Path-free + idempotent (verified across two captures). Suite count unchanged
(snapshot added to an existing example).
- Widened abiCoerceParamType + needsByval to pub (visibility only;
abiCoerceParamTypeEx/materializeByvalArg/verifySizes stay private — move with
callers in sub-step 2). No logic touched.
- Recorded the A7.1 coverage inventory + residual gaps (wasm32 usize->i32 branch,
fn-ptr large-aggregate 1203/1204) in ARCH-SAFETY.md.
Gate: zig build, zig build test, bash tests/run_examples.sh -> 361/0 (no churn
beyond the new 1202 .ir).
Codex review of 95895a3 found 1051 reached neither lambda arm it claimed to
pin: the lambda arrived only as a var_decl initializer, which routes through
checkCleanupNode's `.var_decl` arm -> cleanupReject(lambda) -> early-return
(a lambda literal is not failable), so the `.lambda` stop never ran; and its
accepted-direction `if !err` guard would still pass with flowExpr's lambda
recursion removed.
Scaffolding-only fix (no compiler change):
- 1051: add a bare lambda STATEMENT `() -> !E { failing(); };` in the cleanup
body so checkCleanupNode sees a `.lambda` node directly and stops (the bare
failable inside is accepted; were the arm to recurse it would reject like
1052). Output byte-identical — only the .sx gained the statement.
- 1053-errors-nested-lambda-liveness-reject (exit 1): an E1.8 value-slot read
inside a never-called nested lambda, rejected only because flowExpr recurses
via `.lambda => analyzeFnBody`. Remove that arm and the diagnostic vanishes
-> suite fails. This is the discriminating negative 1051 lacked.
Gate: zig build test, bash tests/run_examples.sh -> 361/0.
Test-first scaffolding for the path-sensitive error-flow pass
(checkErrorFlow/analyzeFnBody/flowWalk/flowIf/checkCleanupBody) before it
moves into src/ir/error_flow.zig. No compiler change — both examples lock
current behavior.
- 1051-errors-cleanup-closure-boundary (accepted): a closure literal inside a
`defer` body is its own function boundary — the E1.7 cleanup rule and the
parser's try/raise ban both stop at the lambda, and E1.8 value-slot liveness
runs per-boundary. Pins checkCleanupNode's `.lambda` stop + flowExpr's
`.lambda` recursion. Constructible since issue 0073 (0310).
- 1052-errors-cleanup-transitive-reject (exit 1): the E1.7 cleanup check is
transitive — bare failables nested in an `if` (both branches), a nested
block, and a `while` body all reject. Pins checkCleanupNode's recursive arms,
distinct from 1049's direct-body case.
No .test.zig/.ir: diagnostic-pass altitude (checkErrorFlow/A2.4 precedent) —
the pass returns no fact object and emits no IR.
Gate: zig build, zig build test, run_examples.sh -> 360/0.
A closure literal declared inside a `defer` body segfaulted the compiler.
Root cause: lowerLambda never opened its own `func_defer_base` window. Every
other function-lowering entry (lowerFunction / monomorphizeFunction /
monomorphizePackFn) saves func_defer_base, sets it to defer_stack.items.len, and
restores it — lowerLambda didn't. So a lambda's `return` drained the ENCLOSING
function's defers; when the defer body itself declared the lambda, draining
re-lowered the lambda, which returned, which drained again → infinite recursion
→ stack-overflow SIGSEGV (the failable variant surfaced one frame out, in
expandCallDefaults→lookupFn reading a clobbered scope).
Fix: lowerLambda now saves func_defer_base + the defer_stack length, sets the
base to the current length (a fresh window), and restores both on exit — so a
lambda's `return` drains only its own defers.
Regression: examples/0310-closures-closure-literal-in-defer.sx — a closure
declared and called inside a `defer`; verifies `body` then `defer closure: 42`
at scope exit (exit 0). Issue 0073 marked RESOLVED; repro promoted from
issues/0073-*.sx.
zig build, zig build test, tests/run_examples.sh (358/0) all green.
Test-first scaffolding ahead of extracting src/ir/error_analysis.zig — no code
change to the convergence targets (convergeInferredErrorSets /
convergeClosureShapeSets / collectErrorSites / collectClosureShapes).
Adds 2 unit tests via the already-pub convergence functions (no new exposure):
- convergeInferredErrorSets transitive/SCC: a `caller :: () -> ! { try raiser(); }`
with no direct raise converges to raiser's {Foo} across the try edge — the
whole-program fixpoint A5.1 must preserve. (Today's E1.4b test only covered a
direct raiser + the empty-set warning.)
- convergeClosureShapeSets: a bare-`!` closure literal `() -> ! { raise error.Bar }`
inside a host fn unions {Bar} into one shape_inferred_sets entry.
Adds 2 .ir snapshots (first .ir for these error forms), vetted clean
(idempotent, path-free, no #run): 1006-errors-inferred-error-sets (inferred-set
error-channel shapes) and 1009-errors-catch (catch lowering). 1004-errors-try
was already pinned.
PLAN-ERR is complete/idle, so the A5 overlap risk is low (the target functions
are stable, not in-flight). The sub-step-2 module will be named
src/ir/error_analysis.zig, NOT errors.zig (src/errors.zig is the DiagnosticList).
zig build, zig build test, tests/run_examples.sh (357/0) all green.
Test-first scaffolding ahead of extracting src/ir/conversions.zig — no code
change to the coercion targets (lowerXX / coerceToType / coerceOrErase /
buildProtocolErasure / tryUserConversion / failable-adapter selection).
Adds 4 .ir snapshots (first .ir for 01xx/09xx/10xx), each captured surgically
via `sx ir | normalize_ir`, path-free, idempotent, and print-free at IR-gen time
(0114-types-build-block-convert was rejected — it prints `--- void / 0 args ---`
+ sx source at IR-gen):
- 0107-types-int-cmp-in-float-ternary numeric int<->float coercion
- 0903-optionals-optional-roundtrip optional wrap/unwrap
- 0904-optionals-any-to-string-optional xx unbox_any + optional
- 1004-errors-try error-channel adapter/coercion
Protocol erasure + user Into are already pinned by the 04xx snapshots
(0400/0413/0414/0416); duplicate-conversion rejection by the 0410/0411/0412
anchors.
Adds 1 unit test via the public surface (no new exposure, mirroring A4.1/A4.2
sub-step 1): optionalOfFlattened — the optional wrap/flatten coercion rule
(T -> ?T; ?T -> ?T, never ??T; contrasted with the non-flattening optionalOf).
The lowerXX/coerceToType/coerceOrErase/buildProtocolErasure decisions are private
+ emission-bound, so their CoercionPlan unit tests land with the extracted module
in sub-step 2.
zig build, zig build test, tests/run_examples.sh (357/0) all green.
The 0524-packs-generic-fn-pack-state-leak example has a #run that prints at
IR-gen time, and tests/run_examples.sh captures `sx ir ... 2>&1`, so its .ir
snapshot was contaminated with #run stdout (`0: len=0` ...) instead of pure IR.
Remove 0524.ir — pack-state isolation (the issue-0048/0050 class) stays guarded
by 0524's existing runtime .stdout/.exit, where a leaked outer pack_arg_types
would corrupt the printed len= sequence.
Replace it with 0513-packs-pack-mixed-comptime.ir, which is print-free at
IR-gen time (clean, idempotent, path-free) and additionally locks the
comptime-value mono-key path (appendComptimeValueMangle): the IR shows
tagged(7,..) vs tagged(9) producing distinct monos
@tagged__ct_7__pack_s64_s64_s64 / @tagged__ct_9__pack.
zig build, zig build test, tests/run_examples.sh (357/0) all green.
Test-first scaffolding ahead of extracting src/ir/generics.zig — no code change
to the refactor targets (buildTypeBindings / mangleGenericName / monomorphize* /
inferGenericReturnType / mangleTypeName).
Adds the first non-FFI generic/pack .ir snapshots (closing the ARCH-SAFETY §3
gap for this phase), each captured surgically via `sx ir | normalize_ir`,
path-free and idempotent:
- 0200-generics-generic generic fn, type-param inference + mono
- 0201-generics-generic-struct generic struct instantiation
- 0507-packs-pack-mono-dedup mono-key dedup (same shape => one mono)
- 0518-packs-pack-value-dispatch pack value dispatch (monomorphizePackFn)
- 0524-packs-generic-fn-pack-state-leak pack-state isolation (issue-0048/0050
class; guards the future scoped-env change)
Adds 2 unit tests via the existing public surface (no new pub exposure,
mirroring the A3.2 sub-step-1 cadence):
- mangleTypeName: pins the mono-key fragment encoding per type shape
(s64 / ptr_X / opt_X / SL_X / mptr_X / AR_n_X / vec_n_X / struct-name / tu_X_Y).
- inferGenericReturnType: explicit type-arg path binds $T and resolves the
-> T return (pair(s64,..) => s64, pair(f64,..) => f64).
The internal substitution/mono-key unit tests (comptime-value mangle,
buildTypeBindings strategies, scoped-env isolation) land with the generics.zig
extraction in sub-step 2, as A3.2's plan-object tests landed with CallPlan.
zig build, zig build test, tests/run_examples.sh (357/0) all green.
globalInitValue's issue-0071 .identifier arm closed the bare-identifier hole,
but .field_access (and every other non-literal expression shape) still fell
through to `else => null`, so a global like `g : s32 = K.x;` was emitted with
no payload and silently zero-initialized (g=0).
Make the `else` emit a diagnostic — "global '<name>' must be initialized by a
compile-time constant" — instead of a null payload, so no unsupported shape can
silently zero. Two arms added alongside:
- `.null_literal => .null_val`: a `*void = null` global was previously a
no-payload zero-init; this preserves the exact LLVMConstNull emission (fixes
3 ffi examples that regressed on the first cut).
- explicit `.enum_literal => null` carve-out: the stdlib's
`OS : OperatingSystem = .unknown;` zero-init is load-bearing for compile-time
`inline if OS == .X`; documented, not folded into a silent fallthrough.
Field-access constant *evaluation* (materializing K.x -> 9) is intentionally
not implemented: a typed struct const like K is not registered in
module_const_map, so it would require new plumbing whose writes are read at
runtime — out of scope. The diagnostic is the issue-sanctioned outcome.
Regression: examples/1118-diagnostics-global-non-const-initializer-rejected.sx
(exit 1). Gate: zig build, zig build test, run_examples.sh -> 356/0.
registerTopLevelGlobal's init_val switch serialized only literal / array-
literal / struct-literal initializers. An identifier initializer
(`K : A : 42; g : A = K;`) fell through to `else => null`, so the global was
emitted with no payload and silently zero-initialized (printed g=0).
Extract the initializer serialization into globalInitValue and add an
.identifier arm that materializes the global's static value from
ProgramIndex.module_const_map (typed module consts are registered in the same
scanDecls pass-2 just before, via registerTypedModuleConst). An identifier
that names no usable constant now emits a diagnostic instead of silently
zeroing — a global has no run site for a dynamic initializer.
Other initializer shapes (enum-literal shorthand, etc.) keep their established
static-lowering behavior; enum-literal globals' zero-init is load-bearing for
`inline if OS == ...` in the stdlib, so it stays out of scope here. This pass
only closes the identifier/module-const hole.
Regression: examples/0134-types-global-init-from-module-const.sx (g=42, exit
42). Gate: zig build, zig build test, run_examples.sh -> 355/0.
Issue 0069's resolveForwardIdentifierAliases fixpoint runs at the END of
scanDecls, but top-level var_decl globals and typed module constants had
their annotations resolved via resolveType(ta) inside the SAME scan loop,
before the fixpoint. So a forward identifier alias (`A :: B; B :: s32;`)
used as a global's type (`g : A = 7;`) was still absent from
type_alias_map: resolveType fabricated an empty-struct stub, and the global
got a type mismatching its initializer at LLVM verification (the typed-const
path `K : A : 42;` silently mistyped the constant instead).
Split scanDecls into two passes: pass 1 registers function/type/alias facts,
then resolveForwardIdentifierAliases converges the aliases, then pass 2
registers var_decl globals (registerTopLevelGlobal) and typed module
constants (registerTypedModuleConst) against the converged alias map.
Globals/typed-consts can't be named in a type position, so deferring them
past type/alias registration is order-safe; the untyped module-const branch
(no annotation to resolve) stays in pass 1.
One incidental IR snapshot reorder (examples/1309: user globals now emit
after foreign-class globals — semantically identical, program still exits 0).
Regression: examples/0133-types-forward-alias-global.sx (forward-alias global
+ typed const). Gate: zig build, zig build test, run_examples.sh -> 354/0.
scanDecls' `.identifier` alias branch registered `A :: B` into
ProgramIndex.type_alias_map only when `B` was already known (in
type_alias_map or the TypeTable). A forward target declared later
(`MyChain :: MyInt; MyInt :: s32;`) was never present during the single
forward scan, so the alias name went unregistered and the A2.4
unknown-type pass — which treats type_alias_map keys as declared types —
flagged its uses as `unknown type 'MyChain'`.
Add a fixpoint post-pass `resolveForwardIdentifierAliases` at the end of
scanDecls that re-resolves identifier-RHS aliases until no progress, after
every top-level name has been seen. A value const is never an `.identifier`
node, and an alias whose target is a value const still misses both lookups,
so issue 0068's value-const rejection is preserved.
Regression: examples/0132-types-forward-type-alias.sx (forward alias +
forward chain). Gate: zig build, zig build test, run_examples.sh -> 353/0.
The A2.4 unknown-type pass (semantic_diagnostics) added EVERY const_decl name to
its declared-type-name set. A value const (`NotAType :: 123`) thus satisfied
reportIfUnknownType, so `v: NotAType` was not flagged; lowering then hit
TypeResolver.resolveNamed's empty-struct-stub fallback and fabricated
`NotAType{}` (the program ran, printing it).
Fix: collectDeclaredTypeNames and harvestScopeDecls now gate the const-name-add
on a new constValueIntroducesType — true only when the value introduces a type
(declarations: struct/enum/union/error; type-expression aliases: type_expr,
pointer/many-pointer/slice/optional/array/function/closure/tuple, parameterized).
`.identifier` / `.call` aliases are intentionally excluded: the scan registers
the type-valued ones into ProgramIndex.type_alias_map / the TypeTable (both
queried separately by the pass), so a value-RHS alias is correctly left out and
flagged, while a type-RHS alias stays covered by the canonical facts.
Regression: examples/1117-diagnostics-value-const-as-type-rejected.sx (exit 1).
Issue-0064 regressions 1111-1116 and the 0115 aliases stay green. Gate: zig
build, zig build test, run_examples 352/0.
`size_of((s32, 1))` treated the tuple literal as a tuple TYPE: for the non-type
element `1` it emitted a `std.debug.print` and substituted `.s64` for that field,
then compiled and printed a bogus size — a silent fabricated type (the forbidden
silent-fallback pattern).
Fix:
- type_bridge.resolveTupleLiteralAsType: a non-type element now yields
`.unresolved` (no `.s64`, no debug print) — it refuses to fabricate a tuple.
type_bridge is stateless, so this is the binding-free backstop.
- New stateful Lowering.resolveTupleLiteralTypeArg validates each element via
isTypeShapedAstNode, emits a user-facing diagnostic at the offending element's
span, and returns `.unresolved`. Wired into resolveTypeArg (size_of/align_of/…)
and the resolveTypeWithBindings name-fallback; type_bridge builds the tuple
only after validation passes.
Regression: examples/1116-diagnostics-tuple-type-nontype-element-rejected.sx
(exit 1 + diagnostic). Valid `(s32, s32)` still works (0115). Gate: zig build,
zig build test, run_examples 351/0.
Closes the two residual silent holes in the unknown-type diagnostic:
- Nested closure / function bodies. The body walk stopped at closure and
nested-fn boundaries, so a typo'd type in a closure's local annotation
silently became a 0-field struct. `walkBodyTypes` now descends control
flow and expressions to re-enter each closure / nested fn via `checkScope`,
which accumulates that scope's generic + value-`Type` params onto the
parent's — so an inner closure still sees the outer function's `$T` (no
false positive) while a genuine unknown is flagged at any nesting depth.
`harvestScopeDecls` collects type-decl names across the whole body
(including nested scopes) up front so locals are never false-flagged.
- Cast targets. `cast(T)` where `T` is a value-`Type` param (no `$`) cast to
a fabricated empty struct silently; it now gets the tailored `$T` hint. An
unknown *literal* cast target already errors via value resolution, so it's
left to that path — no double diagnostic.
Suite: 350 passed, 0 failed. Regressions: examples/1114 (nested-closure
annotation), 1115 (cast value param).
The signature/field check missed body-level type positions: a local
annotation naming a non-existent type flowed through the empty-struct stub
untouched, so `v: Coordnate = 5` silently compiled and ran (the value
dropped) — an invalid program accepted with no diagnostic.
`checkUnknownTypeNames` now also walks each main-file function body
(`checkBodyTypes`): local var/const type annotations — including inside
if / loop / match / push / defer / onfail blocks and decl-value blocks — are
validated with the enclosing function's generic params in scope, and
body-local `T :: struct/enum/union` declarations are collected first
(`collectBodyDeclNames`) so legitimate locals aren't false-flagged. Nested
function/closure bodies are their own scope and are not descended (safe
under-coverage); explicit `cast(T)` already surfaces its own `unresolved`
diagnostic and is left to it.
Regression: examples/1113 (local annotation of a non-existent type, exit 1).
An identifier used in a type position that resolved to nothing fell through
to `type_bridge.resolveTypeName`'s empty-struct-stub fallback, silently
interning a 0-field struct named after the identifier. A value parameter
mistakenly used as a type (`(T: Type, ...) -> T`, missing the `$`) or a
typo'd type name therefore compiled and ran, rendering as `T{}`.
New post-scan diagnostic pass `checkUnknownTypeNames` (lower.zig Pass 1f)
walks every main-file function signature and non-generic struct field type
and rejects any leaf name that is not a primitive, an in-scope generic param
(`$T` / `type_params`), a declared type, or a real (non-stub) registered
type. The load-bearing empty-struct stub is left intact — forward references
and foreign-class opaque types still depend on it during the scan — and the
pass runs before body lowering, so `hasErrors()` halts the build before any
stub reaches codegen.
A value param used as a type gets a tailored hint to write `$T: Type`; a
genuine unknown gets "unknown type 'X'". Imported concrete types are
recognized via the type table, and inline compound spellings (`[:0]u8`),
arbitrary-width ints (`u1`/`u2`), and `$`-introduced generics (`-> $R`) are
exempted to avoid false positives.
Regressions: examples/1111 (tailored hint) + 1112 (typo'd field type).
A value-position match's arms are now lowered with `target_type` set to
the merge's `result_type`, so positive and negated integer literals pick
the same width. Fixes the `PHI node operands are not the same type as the
result` failure for `if n == { case 0: 100; else: -1; }`-style returns.
Regression: examples/0043-basic-match-value-mixed-width.sx.
Gates: zig build, zig build test, run_examples.sh -> 345 passed.
The block-value rework routes value-position `{ … }` through the same
statement parser as every other block, so a destructure decl (and any
statement form) inside a value-bound block now parses, with the trailing
expression as the block's value. The `defer { … }` half was fixed
earlier (634cf9b). Regression: examples/0042-basic-block-value-destructure.sx.
Gates: zig build test, run_examples.sh -> 344 passed.
A block's value is now its last statement ONLY when that statement is a
trailing expression with no `;`. A trailing `;` discards the value,
leaving the block void. This makes value-vs-statement explicit and lets
the compiler reject "this block was supposed to produce a value".
Compiler:
- Parser records `Block.produces_value` (last stmt is a no-`;` trailing
expression) + `Block.discarded_semi` (the `;` that discarded a value),
via `expectSemicolonAfter`. A trailing expression before `}` may now
omit its `;` (previously a parse error). Match-arm and else-arm bodies
are built value-producing regardless of the arm `;` (arms are exempt —
the `;` is an arm terminator).
- Lowering: `lowerBlockValue` / the block-expr path / `inferExprType`
respect `produces_value`. A value-position block that discards its value
is a hard error (`lowerValueBody` for function bodies; the value-context
`.block` path for if/else branches, `catch` bodies, value bindings,
match arms). Pure-failable `-> !` bodies (value rides the error channel)
and a value-if whose branches are void are handled without false errors.
- `defer`/`onfail` cleanup bodies lower as statements (void), so a
trailing `;` there is fine.
Migration (behavior-preserving — output unchanged):
- stdlib + ~210 examples: dropped the trailing `;` on value-position last
expressions. `format` now ends with an explicit `#insert "return
result;"` (it relied on `#insert`-as-block-value, which `;` discards).
- Two `main :: () -> s32` examples that relied on the old silent
default-return got an explicit trailing `0`.
- Rejection snapshots 0412 / 1013 regenerated (their quoted source lines
lost a `;`); the diagnostics themselves are unchanged.
Docs/tests: specs.md "Block values" section; examples 0040 (rules) + 0041
(rejection); 3 parser unit tests. Filed issue 0066 (pre-existing
match-arm negated-literal phi-width quirk, surfaced not caused here).
Gates: zig build, zig build test, run_examples.sh -> 343 passed,
cross_compile.sh -> 7 passed (also refreshed its stale example names).
A braced `defer` body routed through `parseExpr` + a mandatory trailing
`;`, so it parsed the `{ … }` as a block-EXPRESSION whose statement loop
doesn't handle a destructure decl or a `catch`-statement — `defer { v, e
:= f(); … }` and `defer { x() catch e … }` failed with "expected ';'",
and even `defer { stmt; }` needed a spurious trailing semicolon.
Now the `kw_defer` arm parses a braced body with `parseBlock` (the same
path `onfail` uses), so every statement form works; the bare-expression
form (`defer expr;`) is unchanged. `in_defer_body` is still set before
parsing, so the cleanup-body control-flow bans (return/break/continue/
try/raise) and the E1.7 failable-absorption check still fire.
Resolves the `defer` manifestation of issue 0065 (the general
value-block-in-binding-position destructure remains open). Regression:
examples/1050-errors-defer-block-body.sx.
Gates: zig build, zig build test, run_examples.sh -> 341 passed, 0 failed.
A `defer`/`onfail` body runs while the block is already exiting, so a
failable call there has nowhere to propagate its error. The parser
already bans `try`/`raise`/`return`/`break`/`continue` in cleanup bodies
(f9dd965); this adds the remaining sema rule — a bare (un-absorbed)
failable call must be absorbed locally with `catch` or `or <value>`.
Implemented in the shared error-flow pass (`checkCleanupBody` /
`checkCleanupNode` / `cleanupReject` in ir/lower.zig): when the walk hits
a `defer`/`onfail`, it scans the body transitively (through blocks, `if`,
loops, match arms, `catch` handlers; stopping at nested closures) and
flags any still-failable expression. `catch` / `or value` strip the
error channel, so `exprIsFailable` is false for them — only an unhandled
failable trips the check. This completes ERR PLAN E0–E5 plus the two
deferred E1 follow-ups (E1.7 + E1.8).
New regressions: 1048 (catch/or-value absorbed forms compile + run) and
1049 (bare failable in defer and onfail rejected, exit 1).
Filed issue 0065: a braced `defer { … }` / value-block body routes
through `parseExpr` (not `parseBlock` like `onfail`), so it can't parse a
destructure or `catch`-statement inside. Orthogonal to E1.7 — the spec'd
cleanup absorbers (`catch` / `or value`) parse fine in a `defer` body.
Gates: zig build, zig build test, run_examples.sh -> 340 passed, 0 failed.
